Nemotron Puzzle 75B runs on 64GB M2 Max with mlx-lm PR

A contributor added native `nemotron_h_puzzle` support to mlx-lm via PR #1535, enabling the Nemotron Puzzle 75B model to run on a 64GB M2 Max. The work involved fixing blockwise config support, tensor remapping, and aligning NVIDIA's FP32 norm/router behavior.

• 4-bit expert quantization (with 6-bit dense layers and BF16 output head) uses 42.03 GiB checkpoint size and peaks at 49.68 GB memory.
• The 4-bit configuration achieves an average generation speed of 14.27 tok/s, compared to 10.53 tok/s for the 5-bit variant.
• Local task checks scored 24/30 for 4-bit versus 21/30 for 5-bit, while long-context retrieval was 4/5 versus 0/5.
• A bug causing first-layer SSM output discrepancies was resolved by matching NVIDIA's BF16-to-FP32 promotion boundary, improving cosine similarity to 0.999998.
• The BF16 output head is mandatory; quantizing the 131k-vocab lm_head at 4-bit produced repetitive garbage.

The author notes that the 5-bit checkpoint runs near the 64GB memory ceiling, suggesting that higher-memory configurations might yield different results.